Radiomics analysis of panoramic radiographs using machine learning for the detection of peri-implantitis

BACKGROUND: To develop and validate a deep-learning detection model (Mask R-CNN) and a complementary radiomics-based machine-learning analysis for peri-implantitis detection on panoramic radiographs. METHODS: Panoramic radiographs from 144 patients (mean age 57.2 ± 11.7 years) were retrospectively collected. The peri-implantitis regions surrounding the implants of 144 patients were semi-automatically segmented by two dentomaxillofacial radiology residents. A total of 7,045 radiomic features peri-implant were extracted; 6 key features were selected via variance thresholding, SelectKBest, and LASSO regression. A Mask R-CNN (ResNet-50 backbone) was trained (80% train, 20% validation) with data augmentation. Diagnostic performance was assessed by FROC analysis and compared against six machine-learning classifiers. RESULTS: The Mask R-CNN achieved an F1-score of 0.84 (95% CI 0.80–0.88) and AUC of 0.86 (95% CI 0.82–0.90) on the validation set. The best radiomics-based classifier (XGBoost) reached an F1-score of 0.84. Inter-observer ICC for segmentation was 0.97. CONCLUSIONS: Radiomics-enhanced deep learning can reliably detect peri-implantitis on panoramic radiographs. Prospective multicenter validation is warranted before clinical deployment. TRIAL REGISTRATION: Not applicable.